Later attention queries rescued selected early-loop retrieval errors
Source check: Source checked. Not independent scientific confirmation.
Finding: Moving same-item later-loop attention queries into the first loop corrected 89.8% and 96.9% of selected synthetic two-hop errors in two Ouro models. Queries shuffled between items corrected 23.7% and 20.6%, respectively. The result applies to questions answered incorrectly in the first loop and correctly later.
Technology: Ouro-1.4B; Ouro-2.6B; activation transplantation
Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.
- AI contribution
- The authors intervened on pretrained Ouro attention query activations and compared matched late-loop and shuffled donors.
- Test and result
- The primary synthetic test included 600 scenes per model, with 245 and 223 qualifying early-loop errors, respectively. Same-item and shuffled queries were compared at aligned positions.
- Limitation / still unknown
- Natural-question rescue was smaller and did not establish bridge redirection. A positional contribution remains possible. In Ouro-1.4B the rescue changes four-candidate ranking, not greedy generated answers. The intervention supplies a separate query at every layer/head, not one vector; the two model sizes share developmental history.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2610.34639v1 · https://arxiv.org/html/2610.34639v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: linked separately.
Review date: 21 Nov 2026. Review scope: v1 query swap Methods, Table 2 and natural-question caveats.
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 2/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.